fvialibre/edia repository overview

⁠EDIA: Stereotypes and Discrimination in Artificial Intelligence

Language models and word representations obtained with machine learning contain discriminatory stereotypes. Here we present the EDIA project (Stereotypes and Discrimination in Artificial Intelligence). This project aimed to design and evaluate a methodology that allows social scientists and domain experts in Latin America to explore biases and discriminatory stereotypes present in word embeddings (WE) and language models (LM). It also allowed them to define the type of bias to explore and do an intersectional analysis using two binary dimensions (for example, female-male intersected with fat-skinny).

EDIA contains several functions that serve to detect and inspect biases in natural language processing systems based on language models or word embeddings. We have models in Spanish and English to work with and explore biases in different languages ​​at the user's request. Each of the following spaces contains different functions that bring us closer to a particular aspect of the problem of bias and they allow us to understand different but complementary parts of it.

For more information about the project and installation details go to https://github.com/fvialibre/dockerized_edia⁠.

⁠Citation Information

@misc{https://doi.org/10.48550/arxiv.2207.06591,
    doi = {10.48550/ARXIV.2207.06591},
    url = {https://arxiv.org/abs/2207.06591},
    author = {Alemany, Laura Alonso and Benotti, Luciana and González, Lucía and Maina, Hernán and Busaniche, Beatriz and Halvorsen, Alexia and Bordone, Matías and Sánchez, Jorge},
    keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), 
    FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {A tool to overcome technical barriers for bias assessment in human language technologies},
    publisher = {arXiv},
    year = {2022},
    copyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International}
}

⁠License Information

This project is under a MIT license⁠.